About paper question

In the paper, VIT and a general language model, i.e. a language model without reasoning ability, are used to train an MLP. Then, this MLP is used as a bridge between VIT and the language model with reasoning ability to form the main model in the paper. Moreover, this model can achieve good results without the need for an additional SFT. In this step of testing, when the data is input into the network, is its preprocessing done in the same way as the network that trained the MLP or the LLM with reasoning ability? Because when I tested it, it could not infer anything when I processed the data according to the original network that trained the MLP. It did not recognize this format. Looking forward to your reply!

Start by comparing the preprocessing used to train the MLP with the preprocessing used during inference, as described in the issue and the paper. Reproduce the reported failure and determine whether the input format matches the model's expected format; done means inference works or the preprocessing discrepancy is documented.